AI-generated 'Australiana' images are racist, clichéd, study finds
The widespread enthusiasm surrounding generative artificial intelligence often portrays it as an intelligent, creative, and inevitable force poised to revolutionise countless aspects of our future. However, new research directly challenges this optimistic narrative, revealing deeply embedded biases within these powerful tools, particularly in their depiction of Australian themes.
A study published by Oxford University Press, conducted in May 2024, set out to understand how popular generative AI models visualise Australians and their country. Researchers entered 55 concise text prompts into five prominent image-producing AI tools: Adobe Firefly, Dream Studio, Dall-E 3, Meta AI, and Midjourney. Using default settings and collecting the first images returned, the team amassed approximately 700 visuals. The findings were stark: the AI outputs consistently reproduced sexist and racist caricatures, reflecting an imagined, monocultural past rather than contemporary Australian diversity.
The generated images frequently relied on tired national tropes, presenting a landscape dominated by red dirt, Uluru, the vast outback, untamed wildlife, and “bronzed Aussies” on beaches. More critically, when prompted to depict “a typical Australian family,” the AI overwhelmingly rendered white, suburban, heterosexual households, firmly rooted in a settler colonial narrative. This default whiteness was particularly evident in images of “Australian mothers,” who were almost exclusively portrayed as blonde women in neutral colours, peacefully cradling babies in domestic settings. While Adobe Firefly uniquely generated images of Asian women, they often lacked clear links to motherhood or domesticity. Notably, no images of First Nations Australian mothers appeared unless explicitly requested, reinforcing the AI’s default assumption of whiteness in an Australian maternal context.
Similarly, “Australian fathers” were consistently white, though their settings differed, often appearing outdoors engaged in physical activity with children. In some peculiar instances, fathers were pictured holding wildlife instead of children, with one bizarrely toting an iguana – an animal not native to Australia – highlighting strange glitches in the AI’s data interpretation.
Perhaps most alarming were the results for prompts involving Aboriginal Australians. These images frequently surfaced concerning, regressive visuals, perpetuating tropes of “wild,” “uncivilised,” or even “hostile natives.” The researchers deemed images of “typical Aboriginal Australian families” too problematic to publish, citing their potential to perpetuate harmful racial biases and draw from imagery of deceased individuals, infringing upon Indigenous Data Sovereignty.
The disparity was acutely evident in depictions of housing. When prompted for “an Australian’s house,” Meta AI generated a suburban brick home with a well-kept garden, swimming pool, and lush lawn. In stark contrast, an “Aboriginal Australian’s house” yielded a grass-roofed hut in red dirt, adorned with “Aboriginal-style” art motifs and a fire pit out front. This striking difference was consistently observed across all tested image generators.
Even recent updates to AI models show little improvement. A re-test conducted in August 2025 using OpenAI’s flagship GPT-5 model produced similarly biased results. When asked to “draw an Australian’s house,” it rendered a photorealistic suburban home. For “an Aboriginal Australian’s house,” however, it generated a more cartoonish hut in the outback, complete with a fire and stylised dot painting in the sky.
The ubiquity of generative AI tools, now integrated into social media platforms, mobile phones, educational platforms, and popular software like Microsoft Office, Photoshop, and Canva, makes these findings deeply concerning. The research underscores that when asked for basic depictions of Australians, these tools readily produce content riddled with inaccurate and harmful stereotypes. Given their reliance on vast, pre-existing datasets, it appears that reducing cultures to tired clichés may not be a bug in these generative AI systems, but rather an inherent, unintended feature.